Computer Vision News - January 2022

7 Resolution-robust Large Mask Inpainting resolutions. The difference was shown to be even greater with wider holes and surprisingly, it can generalize to high, never seen resolutions! You should have a look at the original paper, especially to see the interesting results of comparisons between the specific architecture and the baselines. An ablation study was also done to examine the influence of different components of the method. Most real-world image editing scenarios require inpainting to work in high-resolution. FFC-based models transfer to higher resolutions significantly thus the hypothesis is made that FFCs are more robust across different scales due to i) image-wide receptive field ii) preserving the low frequencies of the spectrum after scale change, i.e. scale equivariance of 1×1 convolutions in the frequency domain. Showcase of results For thismonthwewill present the results in a gallery format, as wewere lucky the authors provided us with an amazing resource website. Feel free to browser, explore and try to explain based on the theory explained previously! LaMa generalizes with super high resolutions (that includes a resolution more/around 2k). This allows to create examples such as shown in the short videos here: